Você está na página 1de 5

Satisficing, Optimization, and Adaptive Systems

JASON BROWNLEE
Technical Report 070305A
Complex Intelligent Systems Laboratory, Centre for Information Technology Research,
Faculty of Information Communication Technology, Swinburne University of Technology
Melbourne, Australia
jbrownlee@ict.swin.edu.au

Abstract- An aspect of complex adaptive systems is an acting approximately rational, rather economically
observation that they typically operate far from equilibrium and rational in terms of utility. Simon had a problem with the
idea of the traditional economic and rational agent,
 
 

optimality. This provokes investigation into frameworks into


such decision-making processes. Satisficing describes a rational because in reality, agents do not always optimize their
decision making process in economics where deciding agents decision making process. He drew from the field of
psychology to describe a characteristic decision maker .


accept solutions that achieve a minimum level of satisfaction.


This theory differs from previously traditional rational decision- In so doing, he outlined a scenario of rational choice:
making in which the agent seeks to maximise or optimize utility
(1) There are set of alternatives open to choice
from the choices faced. This work investigates some of the impact
this altered perspective on the decision making process has had (2) The relationships that determine payoffs
in economics, game theory, control theory, and evolutionary (satisfaction or attainment) are a function of the
biology. We see it is the latter case that that may be most alternatives that are chosen
interesting to the study of complex adaptive systems and (3) There is preference orderings among payoffs
resultant models, such as those used in biologically inspired (ordinal)
computation.
In this generalised scenario, the decision maker must
Keywords- Satisficing, Adaptation, Optimization, Evolution, apply a strategy. Traditional economic strategies of the
Algorithms, Complex Adaptive Systems time include: the max-min rule, the probabilistic rule,
and the certainty rule. Simon s problem was that these


I. INTRODUCTION strategies required information or computation that no


Consider Complex Adaptive Systems (CAS) [14] as human could possess, and hence were unrealistic.
systems of collectives of adaptive agents whose local Satisficing: The selection of an acceptable or satisfactory solution that





interactions result in dynamical non-linear effects and


meet an agents minimum aspiration-level or threshold , a threshold, under
emergent behaviours. According to Holland [14,15], and
 
 

others [35,36,38], such systems operate far from which solutions are deemed unacceptable . A satisficing solution may or



equilibrium and optimality. This is an intriguing notion may not be an optimal economic solution.
given the ubiquitous adaptationist model of evolutionary Figure 1 - Canonical definition of the term satisficing


biology (a popular example of a CAS) as an




optimization process. That is, organisms are optimized


by evolution for their environment. Satisficing differs from a traditional rational
decision-making process in economics in which the
This work softly explores this theme by providing agent assigns a utility to each choice, and then makes a
background on the decision theory of good enough decision based on the probability that each choice would


solutions known as satisficing and a number of occur maximising utility. In contrast, satisficing
interesting and notable works interpreting this notion in solutions are deemed either satisfactory (above the


the context of game theory, control theory, ecology, and agents aspiration level) or unsatisfactory (below the
search. Satisficing is a notion suitable to describe the agents aspiration level). An agent may decide on the
operating far from optimality description of CAS, and a first, or any satisficing choice, the outcome being that


final discussion considers a few implications of the agent is always satisfied, although the solution may
satisficing in designing future artificial evolutionary and or may not be economically optimal. Such a strategy is
immune system algorithms. simple, computationally tractable, and a rational
strategy, although always selecting a satisficing strategy
II. SATISFICING may not be rational.
The Nobel Prize laureate Herbert A. Simon coined Simons satisficing decision-making and bounded


the term satisficing to describe the selection of a good rationality have had a profound effect in many fields,


enough solution, the selection of a decision that meets a not limited to economics, game theory, management


minimum threshold or aspiration level, the selection of theory, psychology, artificial intelligence and
which occurs in the context of incomplete information or evolutionary biology. Some interesting and notable
limited computation [7,8]. This formalism came out of works follow.
his work on bounded rationality , the idea of agents



Some general works include; a recent collection of

CIS Technical Report 070305A March 2007 page 1 of 5


essays on satisficing decision making [25], satisficing in of many different local and relative considerations.
economics and firms [33], decision theory on deciding A popular quote from Dennett from his work on
when a selected capacity has been learned and overt intention theory [4]:
learning should cease [34], and a philosophical treatment
of satisficing verses optimization [24]. ...poor


old Mother Nature makes do,


opportunistically and short-sightedly exploiting
Much work has been done to extend satisficing and whatever is in hand-until we add: she isn t perfect, but
comparative rationality in game theory by Stirling, et


she does the best she can. Satisficing itself can often be
al. not limited to satisficing in games such as the shown to be the optimal strategy when costs of


prisoners dilemma and battle of the sexes [41]. There is searching are added as a constraint .


a recent book dedicated to the topic of satisficing


games [40] including computer models and algorithms. Richardson considers the question of optimization in


Stirling has also done work on decision-making with the context of evolution ecology [30], and the tradition
regard to appropriate skill selection, as opposed to of considering evolution as maximising adaptiveness of
selecting the single best skill [23], and so called an agent. In his treatment, he shifts the focus from
satisficing equilibria [39]. There has also been work by evolution as an optimizing process, to that of the


others on the evolution of cooperation with regard to decision making component natural selection .


satisficing strategies in [42], and resultant cooperation He contends that optimization models are simpler
given satisficing strategies in mutual interest games [3]. than satisficing models and easier to assess. He further
Satisficing has also proven a useful inspiration for contends that satisficing may be considered
decision making in control systems. In particular is the optimization, although a more complex optimization in
development of the so called satisficing control theory many dimensions. Richardson discusses two points in


for nonlinear systems controllers [18-20]. The satisficing which optimization and satisficing differ: 1. Whether the
threshold was paired with epistemic utility theory as a constraints defining an optimal solution are dynamically


multi-objective approach to accepting and rejecting significant , and 2. Structure of the environment . Here
 

propositions. he is referring to the consideration of a sequential


decision making process of satisficing where unlike
Satisficing controllers have been used in robot soccer optimization, evolution is working with what it has
simulations [13], and so called action selection in robot


based on the organisms history, that alternatives for the


theory [27]. Sen [32] provides a collection of essays on organism are not presented in parallel. He does not
satisficing models of control when dealing with propose that satisficing presents a major shift in thinking
uncertainty. Finally, satisficing has been applied to about evolutionary ecology from optimization, rather a
design-to-criteria [37] as a multiobjective approach to


minor change in perspective.


meeting the needs of clients, and in [17] as a theory to
describe the problem of human designers latching onto Satisficing Evolution: Satisficing provides decision making mechanism in
one design (even if the design is sub-optimal), not which to consider evolution by natural selection not as a process of seeking
considering other designs. maximal adaptiveness of an agent to its environment, but rather sufficient
 


adaptiveness . Such adaptation may or may not be optimal in the




ADAPTATION AND OPTIMIZATION consideration of all the constraints of the agent.


Thus far, satisficing has been superficially Figure 2 - A phrasing of evolution as a satisficing rather than optimizing
demonstrated to be an interesting little pocket of
decision theory, it isn t until it is considered in the Also in the field of ecology, satisficing has been


context of ecology and evolutionary biology that the proposed as a model for predator search (foraging) and
interesting link to adaptive systems becomes apparent. parental behaviour [5]. Complaints of such models,


Bracketing satisficing with Darwinian [evolution] when compared to conventional optimizing foraging
may appear contradictory, for evolutionists sometimes models is that they are considered both difficult to test
talk about survival of the fittest. But, in fact, natural and perhaps a subset of optimizing models. Although
selection only predicts that survivors will be fit enough, testable satisficing foraging models have been proposed
[26]. So called information-gap theory 1 has been


that is, fitter than their loosing competitors; it postulates


satisficing, not optimization


incorporated into satisficing foraging models supporting


experimental observations better than conventional
From [11] (pg 192) quoting Simon and Newell. models.
Adaptation by Darwinian evolution by natural
selection and is often phrased survival of the fittest , as


III. SATISFICING IN SEARCH


the optimization of the best or fittest organisms to Satisficing has influenced considerations in
 

their environment. A well-known problem to ecologists computational search and optimization. This section
and evolutionary biologists is that there is no clear summarises some of the influence satisficing has had in
fitness for natural selection to maximise. As Simon and this field of research.
Newell point out in the above quote, perhaps rather than
the conventional optimizing adaptationist discussion of
evolution, a satisficing perspective may be more 1
Information-Gap Decision Theory: Optimizing robustness to failure or
insightful. Evolution phrased in this way would perhaps opportunity to windfall rather than conventional decision theory that
be survival of the comparatively fit not the most fit,


optimizes utility. The gap refers to what is known, and what needs to be
where evolutionary fitness is measured as the trade-off known.

CIS Technical Report 070305A March 2007 page 2 of 5


Simon, among many other things in his prestigious limited to optimization that limits variability, viability,
career, was a founding contribution to the field of plasticity, flexibility in its drive for an optimum. His
artificial intelligence and heuristics. One of his many approach encourages a minimum standard, robustness.
contributions in this field was his ideas of satisficing Claiming to encourage the adaptability of the structure,
search [9]. This involves problems to which the rather than adapting the structure. No implementation is


solutions are all or nothing , where there is no ordering provided of his conceptions.


between solutions, but there are many ways in which to Channon also investigates the same ideas of what
achieve a solution. Thus constraints are placed on the could be called satisficing in genetic algorithms with the
order in which paths are selected to be searched. Such goal of novelty [1] and open-ended, persistent
problems can be phrased as and-or trees. Some related evolutionary systems [2]. He claims that genetic
and extensions include [6] and [31]. The approach also algorithms perform artificial selection with explicit


been used in learning while doing decision making for




fitness function, what he calls abiotic selection . Such




seeking approximately optimal solutions [28]. systems are unable to perform long-term incremental
Ho [43,44] proposed that satisficing provides an evolution with continual emergence. He investigates
interpretation of optimization of difficult and complex what he calls natural or biotic selection in which there


problems where a good enough solution may suffice, is no explicit specification of what is desired; rather


promoting the use of heuristic and approximate evolution is interpreted as theory of local change and
techniques. That is a solution that is in the top n% of satisficing resulting in continual emergent phenomena.
optimal solution. Another satisficing interpretation in the One cannot help but consider such an emergent
context of search is to that of multiple objective interpretation of evolution as a strong match to complex
optimization, that is the concurrent optimization of adaptive systems theory.
multiple orthogonal objectives, not limited to [10], [12],
and [22]. IV. DISCUSSION
Satisficing in Search Satisficing has had an interesting but perhaps still not
fully exploited impact on search and optimization
Threshold: A threshold is imposed with regard to the satisfaction or theory, and may be an important aspect of complex
suitability of a solution. Once a solution meets such a threshold, the search adaptive systems, and in biologically inspired
may be considered satisfied.
Tradeoffs: More than one consideration (such as fitness or utility) is computation. It is the latter which I will briefly discuss
evaluated when seeking a solution. in this section, in particular in the context of
Comparative: from satisficing evolution, solution utility may be considered evolutionary computation and Artificial Immune
comparatively rather than absolutely. This ordinal ranking may further be Systems (AIS).
reduced to the Boolean status of satisfactory and unsatisfactory solutions.
Presence of Optimality: also from satisficing evolution, unlike optimization The need to consider evolutionary algorithms as
that assumes the presence of a global or locally optimum solution, satisficing complex adaptive systems rather than only as optimizers
makes no such assumptions about the existence of an optimum where one is known, but infrequently heeded [16]. Such algorithms
may or may not exist.
can be used for optimization tasks with what Channon
Figure 3 - A summary of the major interpretations of satisficing in search calls artificial selection (like the artificial selection in


the breeding of horses or dogs), although perhaps the


Juric proposes a so called anti-adaptationist genetic satisficing interpretation of evolution (and perhaps other


algorithm [21] based on satisficing ideas of generating inspirations for CAS and computational models) invites
solutions with minimum threshold rather than optimize. investigation into other aspects of such systems.
He claims that the optimizing adaptive method to search
may impede finding solution. He claims satisficing This work has highlighted some potentially fruitful
almost as an antonym to optimizing adaptation as a


avenues for continued and future research. An important


closer match to biology. He lists four main problems he one is that of open-ended systems resulting in emergent
has with the adaptationist approach, claiming that evolution. The specification of looser objective functions
adaptation is the problematic idea that natural selection or perhaps no objective functions where solutions are
chooses a particular design for a particular reason (best discriminated as satisfactory or unsatisfactory. Such
possible result given the circumstance). systems promote adaptability rather than perform
adaptation, potentially providing future robustness in









Juric s problems with adaptationist GA s











uncertain environments.
• Assumes an optimum exists to which the genotype should be
guided
Does satisficing provide a way to reconcile
• Assumes continued application beyond a minimum acceptance evolutionary algorithms as a model of artificial rather
level seeking the best possible solution than natural selection, to reconcile approximate (less-
• Assumes that schemata that are currently useful will remain useful, than-optimal) optimization?
and that those that are useless will remain useless in the future With regard to the immune system of vertebrates,
schemata theorem satisficing points out the obvious, that the system does
Assume that building blocks that serve a purpose now will

continue to serve the same purpose in the future
not provide a perfect defence. If the system were an
optimal adaptation, then there would be no disease.
Figure 4 - The four main problems Juric has with the adaptationist GA Nesse [29] discusses the mal-adaptations that result from
evolution, highlighting the immune system and disease
His anti-adaptation approach is claimed to allow for as a case study. He questions as to why natural selection


the use of evolved attributes for other purposes. It is not left us vulnerable to disease the evolutionary benefit,


CIS Technical Report 070305A March 2007 page 3 of 5


and why bodies were not designed better. 1991.
[7] Herbert A. Simon, A Behavioral Model of Rational Choice
The typical argument from evolutionary biologists is Quarterly Journal of Economics, vol. 69, pp. 99-118, 1955.
that natural selection is an undirected process whose [8] Herbert A. Simon, Rational choice and the structure of the
influence is weak because of the stochastic nature of environment Psychological Review , vol. 63, pp. 129-138, 1956.
genetic drift and mutations. We should not expect bodies [9] Herbert A. Simon and Joseph B. Kadane , Optimal problem-
to be anywhere near optimal. This argument, while solving search: all-or-none solutions Artificial Intelligence , vol. 6, pp. 235-
good, is insufficient, because while many aspects of say 247, Autumn, 1975.
the human are less than optimal, manu others are [10] Hiroyuki Tamura, Tomohiro Shibata, and Itsuo Hatono,
considered quite close to optimal. He proposes a "Multiobjective combinatorial optimization for performance evaluation by a
metaheuristic satisficing tradeoff method," International conference on
framework of six reasons why we are vulnerable to Advances in production management systems, Berlin, Germany, pp. 490-497,
disease, grouped into three propositions. 2000.
Possible reasons why humans are vulnerable to disease [11] Hunter Crowther-Heyck. Herbert A. Simon: The Bounds Of
Reason In Modern America, John Hopkins University Press, 2005.
(1) Selection is slow (time lag), results in: [12] I. Tamura, T. Shibata, S. Tomiyama, and I. Hatono, "A meta-
(a) A mismatch between design and environment heuristic satisfying tradeoff method for solving multiobjective combinatorial
(b) Competition with a pathogen or other organism optimization problems-with application to flowshop scheduling," Proceedings
(2) Selection cannot solve some problems regardless of how much time it has, of the IEEE International Conference on Systems, Man, and Cybernetics
results in: (IEEE SMC '99 Conference), Tokyo, Japan, pp. 539-544, 1999.
(a) Tradeoffs [13] Jay Packard, Satisficing applied to simulated soccer 2003.
(b) Constraints particular to living organisms (path dependence) Electrical and Computer Engineering, Brigham Young University.
(3) We misunderstand what selection shapes, results in: [14] John H. Holland. Hidden Order: How Adaptation Builds
(a) Traits that increase RS at the cost of disease vulnerability Complexity, USA: Addison Wesley Publishing Company, 1995.
(b) Adverse defences are readily mistaken for diseases
[15] John Henry Holland. Adaptation in Natural and Artificial
Figure 5 - Nesse's six reasons why humans may be vulnerable to disease Systems: An Introductory Analysis with Applications to Biology, Control, and
Artificial Intelligence, USA: MIT Press, 1975.
Clearly, the immune system is satisficing rather than [16] Kenneth A. De Jong, "Genetic Algorithms are NOT Function
optimizing, providing a good solution to detecting and Optimizers," Proceedings of the Second Workshop on Foundations of Genetic
Algorithms (FOGA), Vail, Colorado, USA, pp. 5-17, 1992.
neutralising antigen, not necessarily optimal. These
[17] L. J. Ball, L. Maskill, and T. C. Ormerod, Satisficing in
points raise many questions. More importantly, this line engineering design: causes, consequences and implications for design support
of inquiry provokes interesting lines of research in CAS Automation in Construction, vol. 7, pp. 213-227, Jan, 1998.
and AIS such as: [18] M. A. Goodrich, W. C. Stirling, and R. I. Frost, "A satisficing
fuzzy logic controller," Proceedings of the Fifth IEEE International
What other constraints have evolution considered in Conference on Fuzzy Systems, New Orleans, LA, USA, pp. 272-276, 1996.
its evolution, what has natural selection traded off? What [19] M. A. Goodrich, W. C. Stirling, and R. L. Frost, "A satisficing
would a perfect immune system look like? Is there a approach to intelligent control of nonlinear systems," Proceedings of the 1996
mismatch between the requirements and the IEEE International Symposium on Intelligent Control, Dearborn, MI, USA,
implementation of the vertebrate immune system, and if pp. 248-252, 1996.
so what is it? Finally, what if rather than use artificial [20] M. A. Goodrich, W. C. Stirling, and R. L. Frost, A theory of
selection and simulate the superficial aspects of the satisficing decisions and control IEEE Transactions on Systems, Man and
Cybernetics, Part A, vol. 28, pp. 1083-4427, Nov, 1998.
immune system (such as clonal selection), what if we let
such a defensive system evolve via artificial emergent [21] M. Juric, "An anti-adaptationist approach to genetic algorithms,"
evolution? What would a satisficing interpretation of
Proceedings of the First IEEE Conference on Evolutionary Computation,
Orlando, FL, USA, pp. 619-623, 1994.
clonal selection theory reveal, given the theories [22] M. Sakawa, K. Kato, and T. Shibano, "An interactive fuzzy
selectionist basis was inspired by Darwin s evolutionary


satisficing method for multiobjective multidimensional 0-1 knapsack


theory? problems through genetic algorithms," Proceedings of IEEE International
Conference on Evolutionary Computation, Nagoya, Japan, pp. 243-246,
1996.
ACKNOWLEDGMENTS
[23] Michael A. Goodrich, Wynn C. Stirling, and Erwin R. Boer,
Tim Hendtlass for his patience and for providing Satisficing Revisited Minds and Machines, vol. 10, pp. 79-109, Feb, 2000.
useful feedback on drafts of this paper [24] Michael Byron, Satisficing and Optimality Ethics, vol. 109, pp.
67-93, Oct, 1998.
REFERENCES [25] Michael Byron. Satisficing and Maximizing: Moral Theorists on
[1] A. D. Channon and R. I. Damper, "Evolving novel behaviors via Practical Reason, USA: Cambridge University Press, 2004.
natural selection," Proceedings of Artificial Life VI, Los Angeles, USA, pp. [26] P. Nonacs and L. M. Dill, Is satisficing an alternative to optimal
384-388, 1998. foraging theory Oikos, vol. 24, pp. 371 3751993.
[2] Alastair Channon, Evolutionary emergence: the struggle for [27] Paolo Pirjanian, Satisficing Action Selection Sensor Fusion and
existence in artificial biota 2001. University of Southampton. Decentralized Control in Robotic Systems (SPIE Conference), vol. 3523, pp.
[3] Amit Pazgal, Satisficing leads to cooperation in mutual interests 157-168, 1998.
games International Journal of Game Theory, vol. 26, pp. 439-453, Dec, [28] R. Greiner and P. Orponen, Probably Approximately Optimal
1997. Satisficing Strategies Artificial Intelligence, vol. 82, pp. 21-44, Apr, 1996.
[4] D. Dennett, Intentional Systems in Cognitive Ethology: The [29] Randolph M. Nesse, Maladaptation and Natural Selection The
Panglossian Paradigm Defended The Behaviorial and Brain Sciences, vol. 6, Quarterly Review of Biology, vol. 80, pp. 62 702005.
pp. 343-390, 1983. [30] Robert C. Richardson , Optimization in Evolutionary Ecology
[5] D. Ward, The role of satisficing in foraging theory OIKOS , vol. Philosophy of Science Association (PSA) , vol. 1, pp. 13-21, 1994.
63, pp. 312-317, 1992. [31] Russell Greiner, Ryan Hayward, Magdalena Jankowska, and
[6] Dan Geiger and Jeffrey A. Barnett, Optimal Satisficing Tree Michael Molloy, Finding optimal satisficing strategies for and-or trees
Searches National Conference on Artificial Intelligence, vol. pp. 441-445, Artificial Intelligence, vol. 170, pp. 19-58, Jan, 2006.

CIS Technical Report 070305A March 2007 page 4 of 5


[32] Sandip Sen, "Satisficing Models," Published by The AAAI Press, Agents and Multi-Agent Systems, vol. 5, pp. 305-328, Sep, 2002.
Menlo Park, California, USA, Technical Report SS-98-05, 1998. [40] Wynn C. Stirling. Satisficing Games and Decision Making: with
[33] Sidney G. Winter, Satisficing, Selection, and The Innovating applications to engineering and computer science, USA: Cambridge
Remnant Quarterly Journal of Economics, vol. pp. 237-261, 1971. University Press, 2003.
[34] Sidney G. Winter, The Satisficing Principle in Capability Learning [41] Wynn C. Stirling and Michael A. Goodrich, Satisficing games
Strategic Management Journal, vol. 21, pp. 981-996, 2000. Information Sciences: an International Journal, vol. 114, pp. 255-280, Mar,
[35] Simon A. Levin, Ecosystems and the Biosphere as Complex 1999.
Adaptive Systems Ecosystems, Biomedical and Life Sciences and Earth and [42] Youngse Kim , Satisficing and Optimality in 2 x 2 Common
Environmental Science, vol. 1, pp. 431-436, Sep, 1998. Interest Games Economic Theory, vol. 13, pp. 365-375, Feb, 1999.
[36] Simon A. Levin, Complex adaptive systems: Exploring the [43] Yu-Chi Ho , Heuristics, rules of thumb, and the 80/20
known, the unknown and the unknowable American Mathematical Society, proposition IEEE Transactions on Automatic Control, vol. 39, pp. 1025-
vol. 40, pp. 3-19, 2003. 1027, May, 1994.
[37] Thomas Wagner, Alan Garvey, and Victor Lesser, "Satisficing [44] Yu-Chi Ho, On the numerical solutions of stochastic optimization
evaluation functions: The heart of the new design-to-criteria paradigm," problem IEEE Transactions on Automatic Control, vol. 42, pp. 727-729,
UMASS Department of Computer Science, USA, Technical Report TR-96-82, 1997.
Nov 1996.
[38] W. Brian Arthur. Introduction: Process and Emergence in the
Economy . In: The Economy as an Evolving Complex System II, eds. W.
Brian Arthur, Steven Durlauf, and David A. Lane. Reading, Mass, USA:
Addison-Wesley Pub. Co, 1997.pp. 1-04.
[39] W. C. Stirling, M. A. Goodrich, and D. J. Packard, Satisficing
Equilibria: A Non-Classical Theory of Games and Decisions Autonomous

CIS Technical Report 070305A March 2007 page 5 of 5